Skill-Taste(en)

📅 Last updated: 2026-09-21

⚠️ Star count / version info may change over time; this article reflects the public data at the time of writing

Prerequisite: readers are assumed to know the Agent Skills open standard; this article focuses on taste-skill technical details

🏷️ One-line positioning #

Taste-Skill — gives your AI good taste. stops the AI from generating boring, generic slop

Packages “design taste” into a portable skill, so AI generates frontends that avoid generic AI slop.

📦 Basic info #

🔧 Installation #

Install everything (scans the skills/ directory in the repo):

npx skills add https://github.com/Leonxlnx/taste-skill

Install a single skill (use the name: field from the SKILL.md frontmatter, not the folder name):

npx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"

• v2 is the default; the install name is unchanged, so re-running the command upgrades in place; to lock the original v1, use --skill "design-taste-frontend-v1"
• You can also copy the SKILL.md directly into your project, or paste it into a ChatGPT / Codex conversation

🎛️ Adjustable dials (Settings, taste-skill only) #

Three 1–10 numeric dials at the top of the file:

🧠 v2 core mechanism #

Reads the brief → infers the design language → adjusts the three dials.

🚫 Anti-slop measures #

• An 18-item “AI tells” checklist
• Hard bans (production-tested): section number eyebrow labels, version tags, photo credit captions, decorative hero bars, pills on images, locale bars, scroll cues, decorative dots, divider overuse
• Consistency locks: color / shape / theme
• Layout discipline: hero discipline, bento grid-count rules, italic descender clearance
• Defaults to Tailwind v4; recommends Motion (formerly Framer Motion) for animation

📚 Skills family (10 code + 3 imagegen) #

Implementation skills (produce code) #

Image generation skills (produce reference images only, not code) #

Selection guide #

• Start with taste-skill by default (safest, general-purpose); use taste-skill-v1 for the original behavior; gpt-taste for stricter output
• Fixing an existing codebase → redesign-skill; visual direction already decided → add soft / minimalist / brutalist; agent keeps truncating → output-skill
• If the deliverable is an image → use the imagegen series, then hand the frames to Codex / Cursor / Claude Code
• For image-to-code-skill, remember to state the flow explicitly in the prompt: follow the skill: generate images, then analyze, then code

❓ FAQ #

📁 Research data #

Background research lives in research/: four subdirectories — root-causes/, remediation/, findings/, laziness/ (the basis for the anti-slop rules).

🔗 Reference links #

*Source: taste-skill GitHub README (last commit 2026-09-20)*